Final Report
January 2017
Dossier Lead:
DLR
Contributors:
SERTIT
Project duration:
11/2014 – 12/2016
Project leader
Dr. Sandro Martinis ([email protected])
Authors
Sandro Martinis1, Stephen Clandillon2, Simon Plank1, André Twele1, Claire Huber2, Mathilde Caspard2, Jérôme Maxant2, Wenxi Cao1, Sadri Haouet2, Eva-Maria Fuchs1
1German Aerospace Center (DLR), German Remote Sensing Data Center (DFD),
Oberpfaffenhofen, 82234 Wessling, Germany
2Service Régional de Traitement d'Image et de la Télédétection (SERTIT), Illkirch
Table of Contents
1 INTRODUCTION ... 5
1.1 OBJECTIVES ... 5
1.2 ASAPTERRA PUBLICATION LIST ... 5
2 ASAPTERRA RESEARCH AND DEVELOPMENT ACTIVITIES ... 8
2.1 TECHNICAL NOTE ON FLOOD DETECTION IN RURAL AREAS ... 8
2.1.1 State-of-the-art ... 8
2.1.2 SAR-based flood detection in rural areas ... 17
2.1.2.1 Single temporal flood detection ... 17
2.1.2.2 Change detection ... 31
2.1.2.3 Geohazard Exploitation Platform – analysis of the InSAR Browse Service results ... 41
2.1.3 Optical-based flood detection in rural areas ... 49
2.1.3.1 Very high resolution optical flood mapping ... 49
2.1.3.2 High resolution optical flood mapping (Sentinel-2) ... 57
2.2 REPORT ON AUXILIARY DATA ... 63
2.2.1 Introduction ... 63
2.2.2 Reference Water Masks ... 63
2.2.3 Digital Elevation Data ... 64
2.2.4 DEM-derived indices and land surface characteristics ... 65
2.3 TECHNICAL NOTE ON FLOOD DETECTION IN VEGETATION AREAS ... 70
2.3.1 State-of-the-art ... 70
2.3.2 Time series analysis of multi-frequency SAR amplitude data in Saxony-Anhalt, Germany ... 74
2.3.3 Time series analysis of SAR amplitude and bistatic coherence at Wabash River, USA ... 80
2.3.4 Semi-automated mapping of flooded vegetation using polarimetric Sentinel-1 and ALOS-2 data ... 84
2.4 TECHNICAL NOTE ON FLOOD AND FLOOD TRACE DETECTION IN URBAN AREAS ... 97
2.4.1 State-of-the-art ... 97
2.4.2 Optical HR and VHR based flood detection in urban areas ... 101
2.4.3 Results concerning flood trace detection in urban areas ... 113
2.5 TECHNICAL NOTE ON LANDSLIDE DETECTION ... 115
2.5.1 Optical data-based landslide detection in rural areas ... 115
2.5.1.1 State-of-the-art ... 115
2.5.1.2 Automatic extraction of landslides ... 116
2.5.1.3 Results summary and perspectives ... 124
2.5.2 SAR-based landslide detection in rural areas ... 127
2.5.2.1 State-of-the-art ... 127
2.5.2.2 Texture analysis based landslide detection using polarimetric SAR – a case study in Taiwan ... 132
2.5.2.3 Entropy based landslide mapping by means of VHR post-event polarimetric SAR data ... 136
2.5.2.4 Further concepts and results on landslide detection-based SAR imagery ... 143
2.6 TECHNICAL NOTE ON OPTICAL DATA-BASED BURNT AREA MAPPING ... 146
2.6.1 State-of-the-art ... 146
2.6.1.1 Techniques ... 146
2.6.1.2 Burnt area detection algorithms ... 147
2.6.2 (Semi-)automatic extraction of burnt areas ... 152
2.6.2.1 Fire mapping in Bullsbrook, Australia ... 152
2.6.2.2 Fire mapping without a SWIR channel over Reunion Island, France ... 154
2.6.2.3 Fire mapping with a SWIR channel over Andalusia, Spain ... 160
2.6.2.4 Results summary and perspectives ... 162
2.7 TECHNICAL NOTE ON SAR CHANGE DETECTION-BASED BURNT AREAS MAPPING ... 164
2.7.1 State-of-the-art ... 164
2.7.2 SAR change detection based burnt area mapping using X-band coherence images ... 170
2.7.3 C-band coherence based burnt area mapping using Sentinel-1 SAR imagery... 175
2.7.3.1 Coherence computed within ASAPTERRA ... 175
2.8 TECHNICAL NOTE ON 3D ANALYSIS IN BURNT AREA MAPPING ... 185
2.8.1 State-of-the-art - soil erosion vulnerability ... 185
2.8.2 Indicator of areas vulnerable to soil erosion based on the RUSLE model ... 189
2.9 TECHNICAL NOTE ON FIRE INDUCED VOLUME LOSS ... 199
2.9.1 State-of-the-art - optical extraction of 3D surfaces ... 199
2.9.2 Optical data-based extraction of 3D surfaces ... 201
2.9.3 State-of-the-art – SAR-based extraction of 3D surfaces ... 203
2.9.4 DSM generation from TanDEM-X data ... 207
3 DEMONSTRATION ... 209
3.1 FLOOD DETECTION ... 209
3.1.1 Charter Call 578: Flood in India ... 209
3.1.2 Charter Call 580: Flood in Australia ... 211
3.1.3 Charter Call 580 and EMSR184: Floods in South Eastern Australia ... 214
3.2 FIRE MAPPING DEMONSTRATORS ... 216
3.2.1 EMSR180: Fires on Thassos Island ... 216
3.2.2 Fire Demonstrator, Marseille area, France ... 220
4 FURTHER RESULTS ... 223
4.1 SURFACE CHANGES OF VILLARRICA VOLCANO USING SENTINEL-1A ... 223
1
Introduction
1.1
Objectives
Optical and radar satellite remote sensing have proven to provide essential crisis information in case of natural disasters, humanitarian relief activities and civil security issues in a growing number of cases through mechanisms such as the Copernicus Emergency Management Service (EMS) of the European Commission or the International Charter ‘Space and Major Disasters’.
The aforementioned programs and initiatives make use of satellite-based rapid mapping services aimed at delivering reliable and accurate crisis information after natural hazards.
Although these services are increasingly operational, they need to be continuously updated and improved through research and development (R&D) activities. The principal objective of ASAPTERRA (Advancing SAR and Optical Methods for Rapid Mapping), the ESA-funded R&D project being described here, is to improve, automate and, hence, speed-up geo-information extraction procedures in the context of natural hazards response. This is performed through the development, implementation, testing and validation of novel image processing methods using optical and Synthetic Aperture Radar (SAR) data. The methods are mainly developed based on data of the German radar satellites TerraSAR-X and TanDEM-X, the French satellite missions Pléiades-1A/1B as well as the ESA missions Sentinel-1/2 with the aim to better characterize the potential and limitations of these sensors and their synergy. The resulting algorithms and techniques are evaluated in real case applications during rapid mapping activities.
The project is focussed on three types of natural hazards: floods, landslides and fires.
1.2
ASAPTERRA publication list
The following studies have been published within the ESA-funded project ASAPTERRA and have been used as basis for the content in chapter 2:
1. Martinis, S., Caspard, M., Plank, S., Clandillon, S., Haouet, S., 2017 (submitted): Mapping burn scars, fire severity and soil erosion susceptibility in southern France using multisensoral satellite data. IGARSS 2017, Fort Worth, USA, 23.-28.07.2017.
2. Plank, S., Jüssi, M., Martinis, S., Twele, A., 2017 (submitted): A Combining polarimetric Sentinel-1 and ALOS-2/PALSAR-2 imagery for mapping of flooded vegetation. IGARSS 2017, Fort Worth, USA, 23.-28.07.2017.
3. Cao, W., Plank, S., Martinis, S., 2017 (submitted): Automatic SAR-based Flood detection using hierarchical tile-ranking thresholding and fuzzy logic. IGARSS 2017, Fort Worth, USA, 23.-28.07.2017.
4. Martinis, S., Brcic, R., Plank, S., Tavri, A., Rodriguez Gonzalez, F., 2017 (submitted): The use of the Sentinel-1 InSAR Browse service on ESA’s Geohazards Exploitation Platform for
5. Cao, W., Twele, A., Martinis, S., Plank, S., 2017 (submitted): A three-class change detection methodology for SAR-data based on hypothesis testing and Markov Random Field modelling. International Journal of Remote Sensing.
6. Plank, S., Jüssi, M., Martinis, S., Twele, A, 2017 (submitted): Mapping of flooded vegetation by means of polarimetric Sentinel-1 and ALOS-2/PALSAR-2 imagery. International Journal of Remote Sensing.
7. Martinis, S., Clandillon, S., Twele, A., Huber, C., Plank, S., Maxant, J., Cao, W., Caspard, M., May, S., 2016: Improving the extraction of crisis information in the context of flood, landslide, and fire rapid mapping using SAR and optical remote sensing data. EGU 2016, Vienna, Austria, 17-22 April 2016.
8. Plank, S., Martinis, S., Twele, A., 2016: Combining pre-event optical and post-event polarimetric SAR data for rapid landslide mapping. TerraSAR-X/TanDEM-X Science Team Meeting 2016, Oberpfaffenhofen, Germany, 17-20 October 2016.
9. Huber, C., Clandillon, S., Martinis, S., Twele, A., Plank, S., Maxant, J., Cao, W., Haouet, S., Yésou, H., May, S., 2016: Improving the extraction of crisis information in the context of flood, fire, and landslide rapid mapping using SAR and optical remote sensing data. IGARSS 2016, Beijing, China, 10-15 July 2016.
10. Chow, C., Twele, A., Martinis, S., 2016: An assessment of the ‘Height Above Nearest Drainage’ terrain descriptor for the thematic enhancement of automatic SAR-based flood monitoring services. In: Proceedings of SPIE, SPIE Remote Sensing 2016, 26-29 September 2016, Edinburgh, United Kingdom.
11. Tavri, A. 2016: Flood monitoring based on multi-temporal Sentinel-1 data - a synergistic approach of amplitude data with interferometric coherence. Mater thesis, Technical University of Munich, 69 pages.
12. Twele, A., Cao, S., Plank, S., Martinis, S., 2016a: Sentinel-1 based flood mapping: a fully-automated processing chain. International Journal of Remote Sensing, 37:13, 2990-3004. 13. Plank, S., Twele, A., Martinis, S., 2016: Landslide mapping in vegetated areas using change
detection based on optical and polarimetric SAR data. Remote Sensing, 8, 307, doi:10.3390/rs8040307.
14. Clandillon, S.; Bouillot, L.; Caspard, M.; Haouet, S.; May, S., 2016: Developing improved optical rapid mapping methods by combining Orfeo Tool Box and recent satellite data in the fire mapping domain, ESA Living Planet Symposium 2016, Prague, Czech Republic, 09-13 May 2016.
15. Plank, S., Martinis, S., Twele, A., 2016: Rapid landslide mapping by means of post-event polarimetric SAR imagery, ESA Living Planet Symposium 2016, Prague, Czech Republic, 09-13 May 2016.
16. Twele, A., Martinis, S., Cao, W., Plank, S., 2016b: Automated flood mapping and monitoring using Sentinel-1 data. ESA Living Planet Symposium, Prague, Czech Republic, 09-13 May, 2016.
17. Martinis, S., Rieke, C., Fissmer, B., 2015: Time series analysis of multi-frequency SAR backscatter and bistatic coherence in the context of flood mapping. Multitemp 2015, Annecy, 22-24 July 2015.
18. Plank, S., Hölbling, D., Eisank, C., Friedl, B., Martinis, S. and Twele, A., 2015: Comparing object-based landslide detection methods based on polarimetric SAR and optical satellite imagery – a case study in Taiwan. 7th International Workshop on Science and Applications of SAR Polarimetry and Polarimetric Interferometry, POLInSAR 2015, Frascati, Italy, 27.-30. Jan. 2015.
19. Fissmer, B., 2015: Multitemporal analysis andstatistical evaluation of radar backscatter and bi-Static coherence of flood affected areas. Mater thesis, University of Bochum, 73 pages. 20. Twele, A., Martinis, S. Cao, W., Plank, S., 2015: Inundation mapping using C- and X-band
SAR data: From algorithms to fully-automated flood services. ESA Mapping Water Bodies from Space conference, Frascati, 18-19 March 2015.
21. Caspard, M., Haouet, S., Clandillon, S. (submitted): BURNOUT: a rapid mapping burnt area extraction tool. 14th International Wildland Fire Summit, 31 January 2017, Barcelona, Spain.
2
ASAPTERRA Research and Development Activities
2.1
Technical note on flood detection in rural areas
2.1.1 State-of-the-art Introduction
Every year flood disasters occur in many regions of the globe and cause great losses. Flood extent maps derived from optical or Synthetic Aperture Radar (SAR)-based earth observation (EO) data can be a crucial information source for an effective flood disaster management by supporting humanitarian relief organisations and decision makers in their activities during the time-critical crisis response phase (Voigt et al. 2007). In addition they need a precise monitoring because over- and underestimations can have hard consequences, particularly on people. Besides disaster relief operations, such maps can serve as valuable distributed calibration and validation data for hydraulic models of river flow processes (e.g. Bates et al. 1997, Horritt 2000, Aronica et al. 2002, Hunter 2005, Horritt 2006, Pappenberger et al. 2007, Schumann et al. 2009, Hostache et al. 2009, Matgen et al. 2010) and improve the derivation of spatially accurate hazard maps utilised for spatial planning, flood prevention activities, and insurance risk management (e.g. De Moel et al. 2009).
Flood mapping based on optical data
In case of favourable weather conditions, optical EO data are the preferred information source for flood mapping due to their straightforward interpretability and rich information content. Examples for a successful utilisation of optical satellite data for flood mapping can be found in e.g. Blasco et al. (1992), Smith (1997), Wang et al. (2002), Peinado et al. (2003), Van der Sande (2003), Ahtonen et al. (2004), Brakenridge and Anderson (2005), and Ottinger et al. (2013). A detailed review of optical-based flood mapping is provided by Marcus and Fonstad (2008). However, as flood events often occur during long-lasting periods of persistent cloud cover, a systematic monitoring by optical imaging instruments is rarely feasible.
Visible spectral bands, but particularly near and short-wave infrared wavelengths are the most adapted for water detection. Water reflectance is generally low in the red and infrared spectrum, while higher in the blue and green wavelengths. But this theoretical behaviour of course varies according a water body`s texture (wind), its turbidity and water depth.
Since the 90s’, many studies focused on the potential of the infrared domain and more particularly on the SWIR band for water and soil moisture detection. All concluded that EO satellite data with a SWIR band has an increased sensitivity to water and moisture. This leads to a better mapping of flooded and flood affected areas, even when draw off has started. On a colour composition using a SWIR band, the contrast between water bodies and the other landscape elements is far greater (fig. 2.1.1) (Yésou et al. 2003).
Figure 2.1.1: Contribution of SWIR band in flood detection – Gard Flood, September 2002 (Yésou et al. 2003).
The comparison between water extractions over the Central Alsace flood plain (France) on the same image (SPOT-4) but using two different band combinations - with and without SWIR - has shown that the classification accuracy of the final water surface increased by 7.0% with the SWIR (Clandillon et al. 1999).
Significant achievements have been made in water detection but despite those advancements only few algorithms offer a sufficient reliability for general practical use. Common water classification methods could be organised into few categories: Thematic classification including unsupervised classification (Jenny and McCarthy 2003), Euclidian classifiers, Support Vector Machines (SVM) (Gidudu Anthony 2007), Bayesian classifiers (Alecu Corina 2006, Young-Joon Jeon 2004), Decision Trees (Hu Zhuowei 2007), linear un-mixing, thresholding based on single bands (Leen-Kiat Soh 2005, Yuanzhi Zhang 2003) and two band spectral indices (Zhang Quiwen 2007) as well as visual interpretation. Also combinations of various methods are being used to improve accuracies of water extractions. Techniques based on single bands and indices are commonly used due to their simplicity and time-efficiency.
The spectral characteristics of water can change considerably depending on particle charge (turbidity) plus meteorological conditions (clouds, cloud shadows, wind), landcover/terrain (high vegetation, shadowing) and acquisition factors (sun-glint). To improve mapping potential more complex classification procedures need to be tested and applied. R&D for classification improvement is still an important issue (Brakenridge and Anderson 2006, Michael 2007). The greatest challenge is to achieve sufficient robustness and accuracy for producing reliable results independent from the imagery used.
Flood mapping based on SAR-data
For the purposes of systematic flood mapping and monitoring, the use of the microwave region of the electromagnetic spectrum offers some clear advantages compared to sensors operating in the visible, infrared or thermal range. Being an active monostatic instrument and therefore providing its
SPOT 4 NIR, R, G
SPOT 4 SWIR, NIR, R
own source of illumination in the microwave range, SAR is characterised by nearly all-weather/day-night imaging capabilities as the emitted radar signal is able to penetrate clouds and the imaging process is independent from solar radiation. Over the last decades, spaceborne SAR systems have increasingly been used for flood extent mapping. While past and current SAR satellite and space shuttle radar missions with spatial resolutions of the categories HR2 to MR1 have a proven track-record for large-scale flood and inundation mapping in the X- (SIR-C/X-SAR, SRTM), C- (ERS-1/2 AMI, ENVISAT ASAR, RADARSAT-1, RISAT-1, SIR-C/X-SAR), and L-band domain (SEASAT-1, JERS-1, ALOS PALSAR-1/2, SIR-A/B/C/X-SAR), their capability for deriving flood parameters in complex and small-scale scenarios is limited. Since 2007, the successful launch of the European platforms TerraSAR-X/TanDEM-X and the COSMO-SkyMed constellation (CSK) consisting of four satellites marks a new generation of civil X-band SAR systems suitable for flood mapping purposes. These satellites provide data up to the 0.24 m spatial resolution (TerraSAR-X Staring SpotLight mode), permitting an operational derivation of detailed hydrological parameters from space during rapid mapping activities. The potential of these data has been demonstrated by several studies to support flood emergency situations (e.g. Giustarini et al. 2012, Kuenzer et al. 2013a, Kuenzer et al. 2013b, Martinis et al. 2009, 2011, 2013 and 2014, Martinis and Twele 2010, Mason et al. 2012, Matgen et al. 2011, Pulvirenti et al. 2011 and 2012, Pierdicca et al. 2013, Schumann et al. 2010). The Sentinel-1 satellite mission, operated by the European Space Agency (ESA) in the frame of the European Union Copernicus Programme, a constellation of two polar orbiting C-Band SAR sensors, will enable a systematic large-scale flood monitoring with a spatial resolution of 5 x 20 m in the standard Interferometric Wide (IW) Swath mode and a high temporal resolution of up to six days (Sentinel-1A and 1B) over large parts of the land surface. Sentinel-1 is designed to operate in a pre-programmed conflict-free mode which ensures a consistent long-term data archive for flood mapping purposes (Torres et al. 2012) enabling the implementation of fully-automated flood services (Twele et al. 2015, 2016a and 2106b).
Fully-automated flood services
The number of automatic flood mapping algorithms has significantly increased within the last years, particularly in the SAR domain (e.g. Martinis et al. 2009, Pulvirenti et al. 2011, Matgen et al. 2011), and in most cases a certain amount of user interaction is needed for data pre-processing, the collection and adaptation of auxiliary data as well as the preparation and dissemination of the crisis information to end users. Further, the utility of medium-resolution optical data, such as MODIS, for inundation mapping and monitoring, has been demonstrated in numerous flood events by the Dartmouth Flood Observatory (DFO, http://floodobservatory.colorado.edu/). Building upon this work, NASA’s Goddard’s Office of Applied Science proposes an automated global daily flood and surface water mapping service (http://oas.gsfc.nasa.gov/floodmap/). As one of the first SAR-based services, the Fast Access to Imagery for Rapid Exploitation (FAIRE) service hosted on the ESA’s Grid Processing on Demand system (G-POD, http://gpod.eo.esa.int/) provides automatic SAR pre-processing and change detection capabilities which can be triggered on demand by a user via a
web-interface. The application has been extended with flood mapping capabilities which are based on a comparison of the SAR crisis image with corresponding archive/reference data acquired during normal water level conditions. A prototype automated processing algorithm for medium resolution surface water mapping based on systematic and global-scale ENVISAT ASAR Wide Swath acquisitions has been presented by Westerhoff et al. (2013). A further example is the Fully Automatic Aqua Processing Service (FAAPS) which aims to develop a future operational service delivering NRT flood extent maps generated from ESA satellite data (Schlaffer et al. 2012). A fully automated TerraSAR-X based flood service is presented by Martinis et al. (2015). The processing chain which is automatically triggered after satellite data delivery is comprised of SAR data pre-processing, computation and adaption of global auxiliary data, unsupervised initialisation of the classification as well as post-classification refinement by using a fuzzy logic-based approach. The dissemination of flood maps resulting from this service is performed through an online service (see figure 2.1.2) which can be activated on-demand for emergency response purposes (i.e., when a flood situation evolves). During the course of the project, the processing chain of the TerraSAR-X Flood Service (TFS) has been adapted to Sentinel-1 data (Twele et al. 2015, 2016a and 2016b), which enables a systematic disaster monitoring with high spatial and temporal resolutions. In contrast to the TerraSAR-X Flood Service, this is a major advantage as the time-consuming step of tasking new satellite data can be omitted.
Figure 2.1.2: Web client of DLR’s TerraSAR-X Flood Service (TFS).
Flood mapping in rural areas
due to shadowing effects from buildings as a result of the side-looking viewing geometry of SAR satellite systems. Also floods beneath vegetation layers are difficult to detect due to double bounce scattering resulting in a drastic increase of radar backscatter in such areas.
Some of these challenges demand for very specific solutions with respect to classification methodologies. For this reason, separate work packages are dedicated to flood detection in areas of partially submerged vegetation (see chapter 2.3) and flood detection in urban areas (see chapter 2.4). Nevertheless, also when disregarding these two specific cases, the detection of open flood surfaces in rural areas can be difficult since a number of factors (see tab. 2.1.1) can lead to either an over- or underestimation of the actual flood extent. Although shadowing effects from anthropogenic structures (e.g. buildings) are less dominant in rural areas, also vertical natural features, such as individual trees or forests, can cast shadows which can easily be misclassified as floods. While a skilled image interpreter might be able to use contextual information to separate such features from the flood surface, a semi- or fully-automated algorithm which mainly relies on the grey level of a given pixel is usually unable to differentiate between two features with similar backscattering characteristics. The utilisation of ancillary data (e.g. land cover or digital elevation data) to reduce misclassifications is thus mandatory. In case the shadow effects only affect a small group of pixels (e.g. the radar shadow from individual trees), the implementation of a minimum mapping unit (MMU) as a post-classification step can further help to reduce the number of overestimations in the classification.
Table 2.1.1: Factors leading to misclassification of flooding in SAR data as well as their occurrence and impact on the flood classification result (feature range: high +++, medium ++, low +) (Martinis et al. 2015, modified).
Flood overestimation Flood underestimation
Factor Occurrence/
Impact Factor
Occurrence/ Impact Shadowing effects behind vertical
objects (e.g. vegetation, topography, anthropogenic
structures)
+++
Volume scattering of partially submerged vegetation and water
surfaces completely covered by vegetation
+++
Smooth natural surface features (e.g. sand dunes, salt and clay
pans, bare ground)
+++ Double bounce scattering of
partially submerged vegetation ++ Smooth anthropogenic features
(e.g. streets, airstrips) ++
Anthropogenic features on the
water surface (e.g. ships, debris) +
Heavy rain cells +
Roughening of the water surface by wind, heavy rain or high flow
velocity
+ Layover effects on vertical objects
(e.g. topography urban structures, vegetation)
Apart from radar shadowing, also certain unshaded natural surfaces such as sand dunes, salt clay pans or agricultural fields can exhibit backscatter levels similar to open water surfaces and may lead to an overestimation of the flood extent. Factors which usually occur less frequently are heavy rain cells leading to an overestimation particularly in short-wavelength SAR data (X-band). Although less frequently observed in inland water compared to oceans, strong wind conditions can roughen the water surface, leading to an increased backscatter level in dependence of SAR-polarisation, resulting in an underestimation in the flood extent.
Concerning the accuracy of optical-based flood mapping, a number of factors can lead to the under or over-estimation of flood water surfaces. Meteorological factors such as clouds can reduce the visibility of floods and cloud shadowing can lead to extra erroneous ‘flood’ surfaces. Wind can also perturb the spectral characteristics of water/flood surfaces. Floodwater spectral characteristics can change substantially due to: quantity of suspended particles and particle colour (turbidity), landcover/terrain can affect flood water visibility through high vegetation, terrain or building shadowing, submerged vegetation, and acquisition factors (sun-glint). Furthermore, the delineation of water surfaces can be complex due to confusion with water saturated soils (see tab. 2.1.2).
Table 2.1.2: Factors leading to misclassification of flooding in optical data as well as their occurrence and impact on the classification result (feature range: high +++, medium ++, low +).
Flood overestimation Flood underestimation
Factor Occurrence/
Impact Factor
Occurrence/ Impact Shadowing effects of building,
forest areas or trees +++ Water turbidity +++
Confusion between water and
soil moisture ++ Partially submerged vegetation ++
Similar spectral signature of
water with artificial surfaces +
Sun glint and/or strong wind
effects on water surfaces +
References
1. Ahtonen, P., Euro, M., Hallikainen, M., Solbø, S., Johansen, B., Solheim, I., 2004: SAR and optical based algorithms for estimation of water bodies. Technical Report, FloodMan Project, Helsinki University of Technology, Helsinki, Finland.
2. Alecu Corina, O.S., 2006: Multi-resolution analysis of MODIS and ASTER satellite data for water classification. In Proceedings of the SPIE, the International Society for Optical Engineering.
3. Aronica, G., Bates, P. D., Horritt, M. S., 2002: Assessing the uncertainty in distributed model predictions using observed binary pattern information within GLUE. Hydrol Process, 16, 2001– 2016.
4. Bates, P. D., Horritt, M. S., Smith, C. N., Mason, D. C., 1997: Integrating remote sensing observations of flood hydrology and hydraulic modelling. Hydrol Process, 11, 1777–1795. 5. Blasco F., Bellan, M. F., Chaudury, M. U., 1992: Estimating the extent of floods in Bangladesh
6. Brakenridge, G. R., and Anderson, E., 2005: MODIS-based flood detection, mapping, and measurement: the potential for operational hydrological applications. Proceedings of the NATO Advanced Research Workshop, Baile Felix – Oradea, Romania.
7. Brakenridge, G. R., and Anderson, E., 2006: MODIS-based flood detection, mapping and measurement: the potential for operational hydrological applications. Transboundary Floods: Reducing Risks Through Flood Management, pp. 1-12.
8. Clandillon, S., Meyer, C., Bestault, C., Yesou, H., de Fraipont, P., 1999: SPOT 4 MIR: Evaluation thématique hydrologie/milieux humides. Validation des résultats obtenus sur la plaine d'Alsace au cours des projets pilotes SPOT 4 MIR: Décision d’aide à la Recherche No794/98/CNES/7378 - Rapport final SERTIT - CNES, 57p.
9. De Moel, H., Van Alphen, J., Aerts, J., 2009: Flood maps in Europe, methods, availability and use. Nat Hazard EarthSys, 9, 289-301.
10. Gidudu Anthony, H.G., 2007: Classification of images using support vector machines. Cornell University, arXiv: 0709.3967v1, Library.
11. Giustarini, L., Hostache, R., Matgen, P., Schumann, G., Bates, P. D., Mason, D. C., 2012: A change detection approach to flood mapping in urban areas using TerraSAR-X. IEEE T Geosci Remote 51, 2417-2430.
12. Horritt, M. S., 2000: Calibration of a two-dimensional finite element flood flow model using satellite radar imagery. Water Resour Res, 36, 3279-3291.
13. Horritt, M. S., 2006: A methodology for the validation of uncertain flood inundation models. J Hydrol 326, 153-165.
14. Hostache, R., Matgen, P., Schumann, G., Puech, C., Hoffmann, L., Pfister, L., 2009: Water level estimation and reduction of hydraulic model calibration uncertainties using satellite SAR images of floods. IEEE T Geosci Remote, 47, 431.441.
15. Hu Zhuowei, G. H., 2007: Fast flooding information extraction in emergency response or flood disaster. ISPRS Workshop on Updating Geo-spatial Databases with Imagery & The 5th ISPRS Workshop on DMGISs. Urumchi, Xingjiang.
16. Hunter, N. M., Bates, P. D., Horritt, M. S., De Roo, P. J., Werner, M., 2005: Utility of different data types for flood inundation models within a GLUE framework. Hydrol Earth Syst Sc, 9, 412-430.
17. Jenny, M., McCarthy, T. G., 2003: Flooding patterns of the Okavango wetland in Botswana between 1972 and 2000. AMBIO: A journal of the Human Environment.
18. Kuenzer, C., Huadong, G., Huth, J., Leinenkugel, P., Xinwu, L., Dech, S., 2013b: Flood mapping and flood dynamics of the Mekong Delta: ENVISAT-ASAR-WSM based time series analyses. Remote Sensing 5, 687–715.
19. Kuenzer, C., Guo, H., Schlegel, I., Tuan, V. Q., Li, X., Dech, S., 2013a: Varying scale and capability of Envisat ASAR-WSM, TerraSAR-X Scansar and TerraSAR-X Stripmap data to assess urban flood situations: A case study of the Mekong delta in Can Tho province. Remote Sensing, 5, 5122-5142.
20. Leen-Kiat Soh, C.T., 2005: Segmentation of satellite imagery of natural scenes using data mining. IEEE Transactin On Geoscience and Remote Sensing, p. vol. 37(2) 10861099.
21. Marcus, W. A., Fonstad, M. A., 2008: Optical remote mapping of rivers at sub-meter resolution and watershed extents. Earth Surf. Processes Landforms, 33, 4-24.
22. Martinis, S., Twele, A., Voigt, S., 2009: Towards operational near real-time flood detection using a split-based automatic thresholding procedure on high resolution TerraSAR-X data. Nat Hazard Earth Syst, 9, 303-314.
23. Martinis, S., Twele, A., Voigt, S., 2011: Unsupervised extraction of flood-induced backscatter changes in SAR data using Markov image modeling on irregular graphs. IEEE T Geosci Remote, 49, 251-263.
24. Martinis, S., Twele, A., Strobl, C., Kersten, J., Stein, E., 2013: A multi-scale flood monitoring system based on fully automatic MODIS and TerraSAR-X processing chains. Remote Sensing, 5, 5598-5619.
25. Martinis, S., Twele, A., 2010: A hierarchical spatio-temporal Markov model for improved flood mapping using multi-temporal X-band SAR data. Remote Sensing, 2, 2240-2258.
26. Martinis, S., Kersten, J., Twele, A., 2015: A fully automated TerraSAR-X based flood service. ISPRS J Photogramm.
27. Martinis, S., Kuenzer, C., Twele, A., 2015: Flood studies using Synthetic Aperture Radar data. Remote Sensing of Water Resources, Disasters and Urban Studies, Taylor & Francis, submitted. 28. Mason, D., Davenport, I., Neal, J., Schumann, G., Bates, P., 2012: Near real-time flood
detection in urban areas and rural areas using high resolution synthetic aperture radar images. IEEE Trans. Geosci. Remote Sensing, 50, 3041-3052.
29. Matgen, P., Hostache, R., Schumann, G., Pfister, L., Hoffman, L., Svanije, H. H. G., 2011: Towards an automated SAR based flood monitoring system: Lessons learned from two case studies. Phys Chem Earth 36, 241-252.
30. Matgen, P., Montanari, M., Hostache, R., Pfister, L., Hoffmann, L., Plaza, D., Pauwels, V. R. N., De Lannoy, G. J. M., De Keyser, R., Savenije, H. H. G., 2010: Towards the sequential assimilation of SAR-derived water stages into hydraulic models using the particle filter: proof of concept. Hydrol Earth Syst Sc, 14, 1773-1785.
31. Michael, 2007: Multitemporal Remote Sensing for mapping and monitoring Floods an approach involves validation of the KAFRIBA. Master Thesis.
32. Ottinger, M., Kuenzer, C., Liu, G., Wang, S., Dech, S., 2013: Monitoring land cover dynamics in the Yellow River Delta from 1995 to 2010 based on Landsat 5 TM. Appl Geogr, 44, 53-68. 33. Pappenberger, F., Frodsham, K., Beven, K., Romanowicz, R., Matgen, P., 2007: Fuzzy set
approach to calibrating distributed flood inundation models using remote sensing observations. Hydrol Earth Syst Sc, 11, 739-752.
34. Peinado, O., Kuenzer, C., Voigt, S., Reinartz, P., Mehl, H., 2003: Fernerkundung und GIS im Katastrophenmangement – die Elbe Flut 2003: Angewandte Geographische Informationsverarbeitung XV. Beitraege zum AGIT-Symposium Salzburg 2003, Heidelberg, Wichmann, 342-348.
35. Pierdicca, N., Pulvirenti, L., Chini, M., Guerriero, L., Candela, L., 2013: Observing floods from space: Experience gained from COSMO-SkyMed observations. Acta Astronaut, 84, 122-133. 36. Pulvirenti, L., Chini, M., Marzano, F. S., Pierdicca, N., Mori, S., Guerriero, L., Boni, G.,
Candela, L., 2012: Detection of floods and heavy rain using Cosmo-SkyMed data: The event in Northwestern Italy of November 2011. Proc. of IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2012), Munich, Germany, 22-27 July 2012, 3026-3029.
37. Pulvirenti, L., Pierdicca, N., Chini, M., Guerriero, L., 2011: An algorithm for operational flood mapping from Synthetic Aperture Radar (SAR) data using fuzzy logic. Nat Hazard Earth Sys, 11, 529-540.
38. Pulvirenti, L., Pierdicca, N., Chini, M., Guerriero, L., 2013: Monitoring flood evolution in vegetated areas using COSMO-SkyMed data: The Tuscany 2009 case study. IEEE J Sel Top Appl, 99, 1-10.
39. Schlaffer, S., Hollaus, M., Wagner, W., Matgen, P., 2012: Flood delineation from synthetic aperture radar data with the help of a priori knowledge from historical acquisitions and digital elevation models in support of near-real-time flood mapping. In: Proceedings of SPIE, Earth
Resources and Environmental Remote Sensing/GIS Applications III, Edinburgh, 853813-1 - 853813-9.
40. Schumann, G., di Baldassarre, G. D., Paul, P. D., 2009: The utility of spaceborne radar to render flood inundation maps based on multialgorithm ensembles. IEEE T Geosci Remote, 47, 2801-2807.
41. Schumann, G., di Baldassarre, G. D., Alsdorf, D., Bates, P. D., 2010: Near real-time flood wave approximation on large rivers from space: Application to the River Po, Italy. Water Resour Res, 46, 1–8.
42. Smith, L. C., 1997: Satellite remote sensing of river inundation area, stage, and discharge: A review. Hydrol Process, 11, 1427-1439.
43. Torres, R., Snoeij, P., Geudtner, D., Bibby, D., Davidson, M., Attema, E., Potin, P., Rommen, B., Floury, N., et al., 2012: GMES Sentinel-1 mission, Remote Sens Environ, 120, 9-24.
44. Twele, A., Martinis, S., Cao, W., Plank, S., 2015: Inundation mapping using C- and X-band SAR data: From algorithms to fully-automated flood services. Mapping Water Bodies from Space (MWBS 2015), 18-19 March 2015, Frascati, Italy, 46-47.
45. Twele, A., Cao, S., Plank, S., Martinis, S., 2016a: Sentinel-1 based flood mapping: a fully-automated processing chain. International Journal of Remote Sensing, 37:13, 2990-3004.
46. Twele, A., Martinis, S., Cao, W., Plank, S., 2016b: Automated flood mapping and monitoring using Sentinel-1 data. ESA Living Planet Symposium, Prague, Czech Republic, 09-13 May, 2016.
47. Van der Sande, C. J., de Jong, S. M., De Roo, A. P. J., 2003: A segmentation and classification approach of IKONOS-2 imagery for land cover mapping to assist flood risk and flood damage assessment. Int J Appl Earth Obs, 4, 217-229.
48. Voigt, S., Kemper, T., Riedlinger, T., Kiefl, R., Scholte, K., Mehl, H., 2007: Satellite image analysis for disaster and crisis-management support. IEEE T Geosci Remote, 45, 1520-1528. 49. Wang, Y., Colby, J. D., Mulcahy, K. A., 2002: An efficient method for mapping flood extent in
a coastal flood plain using Landsat TM and DEM data. Int J Remote Sens, 23, 3681-3696. 50. Westerhoff, R. S., Kleuskens, M. P. H., Winsemius, H. C., Huizinga, H. J., Brakenridge, G. R.,
Bishop, C., 2013: Automated global water mapping based on wide-swath orbital synthetic-aperture radar. Hydrol Earth Syst Sc, 17, 651-663.
51. Yesou, H., Clandillon, S., Allenbach, B., Bestault, C., de Fraipont, P., 2003: Exploitation de données SPOT 5, SPOT 4 et SPOT 2 pour la cartographie des zones affectées par les inondations de type Cévenols les 9-10 septembre 2002 dans le département du Gard, SIRNAT, Les Journées pour la Prévention des Risques, 29-30 January 2003, BRGM, Orléans, 6 p.
52. Young-Joon Jeon, J. G. C. I., 2004: A study on supervised classification of remote sensing satellite image by bayesian algorithm using average fuzzy intracluster distance. IWCIA, pp. 597-606.
53. Yuanzhi Zhang, J. T., 2003: Water quality retrievals from combined Landsat TM data and ERS-2 SAR data in the Gulf of Finland. IEEE Transacrtion on GEOSCIENCE AND REMOTE SENSING, pp. 0196-2892.
54. Zhang Quiwen, W. C., 2007: Automatic extraction of water body based on EOS/MODIS remotely sensed imagery. In Proceedings of the SPIE, pp. Volume 6786, 678642.
2.1.2 SAR-based flood detection in rural areas 2.1.2.1 Single temporal flood detection
Since the establishment of DLR’s ZKI (Center for Satellite-Based Crisis Information), the development of EO-based methodologies for the rapid mapping of flood situations in rural areas has been of major concern. This is can be especially contributed to the fact that inundations constitute the majority of all ZKI-activations as well as activations of the International Charter ‘Space and Major Disasters’.
These requirements have led to the development of dedicated SAR-based flood mapping tools (“AFD” and “RaMaFlood”) which have been utilised within numerous rapid mapping activities of flood situations since 2008. The core of these tools is an automatic tile-based thresholding approach (Martinis et al. 2009, Martinis and Twele 2010, Martinis et al. 2011) which allows separating inundated areas from land-areas without any user interaction.
Since 2012, the SAR-based flood detection algorithm has been substantially extended and refined in robustness and transferability to guarantee high classification accuracy under different environmental conditions and sensor configurations with the ultimate goal to allow its implementation in an automatic processing chain for TerraSAR-X data (Martinis et al. 2015). The processing chain including TerraSAR-X data pre-processing, computation and adaption of global auxiliary data, unsupervised initialisation of the classification as well as post-classification refinement by using a fuzzy logic-based approach is automatically triggered after new TerraSAR-X data is available on the delivery server. The dissemination of flood maps resulting from the service is performed through a dedicated web client (see fig. 2.1.2). With respect to accuracy and computational effort, experiments performed on a data set of ~400 different TerraSAR-X scenes acquired during flooding all over the world with different sensor configurations confirmed the robustness and effectiveness of the flood mapping service. Currently, the flood mapping service is activated on-demand in case of emergencies (e.g. in the framework of the ‘International Charter Space and Major Disasters’) by tasking suitable TerraSAR-X new acquisitions
The processing chain of the TerraSAR-X flood service (Martinis et al. 2015) has recently been adapted to the new ESA’s C-band SAR mission Sentinel-1 (Twele et al. 2015, 2016b). In the frame of ASAPTERRA, the current prototype has further being improved in robustness and its accuracy has been assessed based on several test sites (Twele at al. 2016a). Both thematic processors (TerraSAR-X and Sentinel-1 based) have further been enhanced through the integration of the “Height above nearest drainage” (HAND) index (Rennó et al. 2008) which helps to reduce water look-alikes depending on the hydrologic-topographic setting. The HAND index will be calculated near-globally based on elevation and drainage direction information provided by the Hydrosheds mapping product (see chapter 2.2.4).
The processing steps of the Sentinel-1 processing chain can be divided into six main parts (see fig. 2.1.3):
1) Data ingestion,
2) Data preprocessing including geometric correction and radiometric calibration, 3) Initial classification using automatic thresholding,
4) Fuzzy logic-based classification refinement,
5) Final classification including reference water masks and the HAND-index, 6) and the dissemination of the results.
Figure 1.1.3: Processing chain of the Sentinel-1 based flood service (prototype).
The processing chain is activated through a Python script which routinely polls the data delivery server (e.g. FTP-site) for new data. In order to allow an immediate processing, data streams from the Collaborative Ground Segment and ESA Sentinel Data Hub can directly be ingested into the processing chain. In an experimental study, a near real-time (NRT) data access to Sentinel-1 data through DFD’s ground station in Neustrelitz is currently being implemented to allow an even faster data access for scenes acquired over European territories. Once new data are found, the data are downloaded to the local file system and the thematic processor is executed. After unzipping the data, the folder structure is searched for files relevant for the further processing, namely Sentinel-1 data in GeoTIFF-format and XML metadata used for radiometric calibration. During the preprocessing step, a range doppler terrain correction of Sentinel-1 data and radiometric calibration
DATA INGESTION GEOMETRIC & RADIOMETRIC
CALIBRATION THRESHOLDINGAUTOMATIC
FUZZY LOGIC-BASED
REFINEMENT CLASSIFICATIONFINAL
DISSEMINATION
Sentinel-1 GRD Data
(.tiff) Meta Data
(.xml) (Range Doppler Eq.) S1TBX GPT
Tiles Selection Initial Classification Global Threshold NRCS Image Slope SRTM Fuzzy Image Fuzzy Value
Calculation GrowingRegion
Reference Water Standing Water Non Water Flood Water Sentinel-1 Delivery Server FTP Pull HAND Geoserver WFS WMS Website/ Webclient
to sigma naught (dB) is performed using the graph processing tool (GPT) of the ESA Sentinel-1 Toolbox (S1TBX), which allows an execution of this processing step within a fully automated processing chain. The SRTM-tiles corresponding to a given Sentinel-1 scene are automatically downloaded by the S1TBX and transformed from sensor coordinate system (Ground Range Detected - GRD) to geographical coordinates (lat/lon, WGS84). After data pre-processing using the S1TBX, the calibrated data is filtered to reduce the SAR-inherent speckle noise. The result of the pre-processing is a reprojected, radiometrically calibrated and rescaled normalised radar cross section (NRCS) image.
The subsequent classification step can be subdivided into three parts: a) a tile-based initial classification using an automatic thresholding algorithm, b) a fuzzy-logic based post-classification, and c) the final classification including reference water masks and the exclude layer derived from the HAND-index. The tile-based initial classification divides the SAR image into a set of tiles, from which a maximum of five tiles are selected to derive the global optimal threshold for flood detection. This result is then improved through a fuzzy-logic based post-classification according to fuzzy values calculated from backscatter, DEM, slope and the size of individual flood objects. The elements of the fuzzy set are defined by standard S and Z membership functions (Pal and Rosenfeld 1988), which express the degree of an element's membership 𝒎𝒎𝒇𝒇 to the class water. The fuzzy threshold values of each element are either determined according to statistical computations or are set empirically. The average of the individual membership degrees is computed for each pixel to combine all fuzzy elements into one composite fuzzy set. Subsequently, the flood mask is derived through a threshold defuzzification step, which transforms each image element with a membership degree > 0.6 into a discrete thematic class. Subsequently, the exclusion mask based on the HAND-index (see chapter 2.2.4) is applied to mask out potential misclassifications in all areas situated ≥15 m above the drainage network. With the aid of the reference water mask (see chapter 2.2.2), it is then possible to classify the SAR image into three types of pixels: a) flood, b) non-flood, and c) standing water.
The output is stored in the local file system and the results can be automatically disseminated via a web-based user interface. While this user interface still needs to be developed for Sentinel-1 data, a comparable interface already exists for the MODIS and TerraSAR-X flood processing chains. In the following sections, first results based on a prototype of the Sentinel-1 based processing chain, including a cross-comparison to the TerraSAR-X flood service, are presented.
Test site: Floods in Malawi/Mozambique, January 2015
The first test site comprises a flood situation which occurred in the border region between Southern Malawi and Northern Mozambique at the end of January 2015. The flood situation was captured by a Sentinel-1 scene acquired on 28 January 2015 in IW mode and VV-polarisation. A subset of the respective scene is shown in figure 2.1.4.
there is certain variability in the flood surface with a backscatter coefficient ranging from -25 dB up to -15 dB. Potential sources for misclassifications include fairly bare soils at the borders of the main flood plain (backscatter coefficient ranging from -10 dB up to -16dB) and a small number of areas with topographically-induced shadow. The scene was automatically classified with the prototypic Sentinel-1 flood processing chain in less than 27 minutes on an Intel Xeon CPU with 3.4 GHz and 32 GB of RAM running Windows 7 (64 bit). The fuzzy threshold was automatically derived between a = -19 dB and b = -15.34 dB. The classification result is shown in figure 2.1.5, including a visualisation of the HAND-based exclude layer in green color.
Figure 2.1.4: Sentinel-1 scene subset of the Malawi/Mozambique 2015 floods, mode: IW, acquisition date: 2015-01-18, polarisation: VV.
Figure 2.1.5: Classification result of the automatic Sentinel-1 based processing chain. Reference water mask: blue, flood extent: light blue, HAND-based exclude layer: green.
A qualitative analysis including a comparison of the original SAR-scene with the derived classification indicates a very satisfying accuracy of the flood extent, with only a low number of misclassifications. A majority of potential misclassifications have been effectively reduced through a combination of different fuzzy sets and the HAND-based exclude layer.
Test site: Floods in the Balkans, May 2014
In order to cross-compare the performance and classification accuracy of the Sentinel-1 flood processor prototype to the existing TerraSAR-X based processing chain, a test site located in the Balkans at the border region between Serbia and Bosnia Herzegovina has been selected. Located at the confluence between the Sava and Drina River, the terrain is relatively flat with elevations ranging from 74 up to 90m. The land cover is dominated by agriculture and a forested area is located in the north-western part of the test site. On 14 May 2014, a heavy rainfall event caused torrential floods across the region, destroying bridges and infrastructure, and leading to numerous landslides.
The satellite data analysed for this flood event comprise a TerraSAR-X Stripmap scene acquired on 21 May 2015 in HH-polarisation and a Sentinel-1 Stripmap (S3-beam) scene acquired on 24 May 2015 in dual-polarisation (HH/HV). Due to the superiority of HH- over HV-polarisation for flood mapping and better comparability to the TerraSAR-X scene, the Sentinel-1 analysis was based on the HH-polarised channel. Although the time-difference between both scenes is three days, the flood conditions remained very stable during this time frame (see upper row of fig. 2.1.6 for a visual comparison). Due to HH-polarisation, which is less sensitive to wind-induced roughening of the water surface, the flooded areas can easily be discerned from non-flooded regions in both X- and C-band data due to their low backscatter signature and further show a relatively low variability of backscatter levels. Major differences in the backscatter levels between X- and C-band data can be observed in agricultural areas, where the backscatter level of Sentinel-1 data is significantly lower. Such differences are potentially caused by the higher penetration depth of C-band data with respect to vegetation and soils and a higher amount of volume scattering in vegetated areas compared to a stronger canopy scattering of X-band data. As an additional factor and when presuming wet soil conditions due to strong rainfall events, the amplitude between the differing penetration depths of both radar wavelengths is higher than in soils with a lower moisture content (Behari 2005).
Figure 2.1.6 (lower row) shows a comparison of the classification results calculated by the TerraSAR-X and Sentinel-1 flood processors for scenes acquired during the Balkans floods in May 2014. A first qualitative analysis including a cross-comparison between Sentinel-1 and TerraSAR-X and a comparison of the original SAR-scene with derived classification indicates a very satisfying accuracy of the flood extent. Although the backscatter level of agricultural areas is relatively low in Sentinel-1 data and less impact of DEM-based classification refinement due to comparably flat terrain characteristics, no major misclassifications (i.e. “false alarms”) can be observed.
Figure 2.1.6: Floods in the Balkans 2014. Upper row (left to right): TerraSAR-X Stripmap scene (21/05/2014), Sentinel-1 Stripmap S3 (24/05/2014); lower row: Automatic classification results from the TerraSAR-X (left) and Sentinel-1 (right) flood processors.
Validation site: Floods in Greece/Turkey, March 2015
Introduction
For a quantitative accuracy assessment of the Sentinel-1 based flood processor, a validation site at the border between Greece and Turkey was chosen. The validation data consists of two subsets of a Sentinel-1 GRD VV/VH scene (ascending, relative orbit: 29) acquired on 12 March 2015. Each subset comprises an area of approx. 5000 x 3800m at River Evros, where a long-lasting flood situation which started in February 2015 caused widespread flooding of farmland (fig. 2.17).
The thematic accuracy has been assessed separately for both standard polarisation configurations (VV and VH) which are available using routinely acquired IW mode Sentinel-1 data. A reference water mask was generated by visual interpretation and manual digitalisation of a pan-sharpened WorldView-2 scene of 0.5m spatial resolution which was acquired on 11 March 2015 (see fig. 2.1.10a and 2.1.11a). Due to a time difference of ~31 hours between the Sentinel-1 and WorldView-2 data set, stable flood conditions have been ascertained using consecutive SAR and optical satellite acquisitions. No significant change in water levels could be observed.
Figure 2.1.7: Overview of validation site locations. Background: Sentinel-1 scene acquired 12 March 2015, VV polarisation. HAND-EM: Exclusion mask derived from the “Height above nearest drainage” (HAND) index (Source: Twele et al. 2016b).
Accuracy assessment
The performance of the proposed algorithm is evaluated by comparing the classification results of the automatic Sentinel-1 based processing chain to the manually derived water extent of the WorldView-2 data set. Producer’s (PA), user’s (UA) and overall (OA) accuracy as well as the KHAT coefficient have been calculated for the classes “open water” and “other”.
In table 2.1.3 and figure 2.1.8, the classification accuracies for the first validation site (“A”) are listed. While no major differences between the classification accuracy of VV and VH polarisation can be observed, VV-polarised data consistently reached higher user’s, producer’s and overall accuracies. The KHAT coefficient of 0.91 for VV in contrast to 0.88 for VH polarisation confirms this observation.
Figure 2.1.8: Classification accuracies for validation site A.
Table 2.1.3: Classification accuracies for validation site A (Source: Twele et al. 2016b).
Polarisation OA PA Open Water UA Open Water PA Other UA Other KHAT
VV 94.69% 91.39% 99.33% 99.43% 92.49% 0.91
VH 93.98% 88.26% 99.20% 99.34% 90.03% 0.88
With respect to OAs and KHAT coefficients, the results obtained for validation site B (see tab. 2.1.4 and fig. 2.1.9) are fairly consistent to validation site A with VV polarisation performing slightly better than VH polarisation. However, UAs and PAs obtained for the individual classes (“open water” and “other”) show a larger variability. While VH polarisation offers the highest PA and VV polarisation the highest UA for the open water class, the PA for the remaining class (“other”) is over 10.0% higher for VV compared to VH polarisation. In contrast, the UA of VH is over 6.0% higher than that of VV polarisation.
Figure 2.1.9: Classification accuracies for validation site B.
Table 2.1.4: Classification accuracies for validation site B (Source: Twele et al. 2016b).
Polarisation OA PA Open Water UA Open Water PA Other UA Other KHAT
VV 96.08% 95.15% 99.16% 98.18% 89.99% 0.91
Figure 2.1.10: Subsets for validation site A: a) WorldView-2 scene acquired on 11 March 2015, band combination 5-3-2 (natural colour). Overlay: reference water mask b) Sentinel-1 scene acquired on 12 March 2015 in VV polarisation and c) in VH polarisation d) classification results for VV-polarised and e) for VH-polarised data (Source: Twele et al. 2016b).
Figure 2.1.11: Subsets for validation site B: a) WorldView-2 scene acquired 11 March 2015, band combination 5-3-2 (natural colour). Overlay: reference water mask b) Sentinel-1 scene acquired 12 March 2015 in VV polarisation and c) in VH polarisation d) classification results for VV-polarised
Discussion
Regardless of validation site and polarisation configuration, the results obtained can be regarded as very satisfying with overall accuracies between 93.98% and 96.08% and KHAT coefficients between 0.88 and 0.91. Sources of misclassifications can be best identified when comparing the backscatter behaviour of both polarisations (fig. 2.1.10b/c and 2.1.11b/c with the corresponding classification results (fig. 2.1.10d/e and 2.1.11d/e).
In the upper right part of figure 2.1.10, thin lines of vegetation within open water areas can be perceived. However, these lines are much wider in the SAR image compared to the optical WorldView-2 scene. Due to geometric distortions such as foreshortening and layover resulting from the side-looking geometry of the sensor, the SAR-derived flood extent is commonly underestimated in these areas. In addition, when comparing the VV and VH polarisation images, a higher contrast between open water areas and the thin lines of vegetation can be observed for VV polarisation. This is due to double bounce effects, i.e. strong signal return of SAR waves, scattered at the water surface towards e.g. partially submerged tree trunks and then directly back to the SAR sensor. Within this scattering mechanism, a great percentage of the backscattered waves show preserved polarisation (i.e., V-polarised wave transmitted returns as V-polarised signal). Therefore, VV shows a higher signal response in double bounce conditions than VH.
The channel at the peninsula (right part of fig. 2.1.10) shows a higher rate of false alarms in VH polarisation (fig. 2.1.10 e) compared to VV polarisation (fig. 2.1.10d). Also when comparing the two corresponding SAR intensity images (fig. 2.1.10b/c), it is much easier to visually separate the channel from surrounding vegetation using VV polarisation instead of VH polarisation. According to Lee and Pottier (2009), VH polarisation is induced by volume scattering. Therefore, VH polarisation shows a higher signal return at the tree crowns which are flipped towards the channel due to the side-looking SAR geometry. In contrast to this, VV polarisation is characterised by a higher penetration depth into the tree crowns, i.e. VV is mainly influenced by double bouncing between tree trunks and the water surface and not by multiple scattering inside the tree crown as VH polarisation.
A further observation is the larger backscatter variability of VH-polarised data in vegetated areas. Within the land areas of figure 2.1.10, forest areas are much more highlighted in VH- compared to VV-polarised data. Contrarily, agricultural farmland shows comparably low backscatter in VH polarisation. The higher sensitivity of VH polarisation to volume scattering (an indication for vegetated areas) is the reason for its higher contrast within the land part compared to VV. The stronger contrast between forested areas and farmland in VH polarisation increases the risk of water lookalike areas. Due to the application of the HAND-EM (see fig. 2.1.7), lookalike areas situated in higher altitudes do not constitute a source of misclassification in VH-polarised data. In lower altitude areas however, certain more sparsely vegetated land cover types can exhibit similar backscatter profiles like open water surfaces. An example of such misclassifications is visible in figure 2.1.11c/e where a relatively high false alarm rate for open water can be observed.
When comparing the backscatter behaviour of VV and VH polarisation, it is further important to note that the Sentinel-1 scene used for the underlying study has been recorded under calm wind conditions of approx. 1.5m/s. Therefore, open water areas are mainly characterised by specular reflection of the SAR signal which results in relatively low backscatter values. The amount of specular reflection decreases when the water surface is roughened by wind, leading to an increased backscatter level and lower contrast to land areas. In a study to compare quad-polarised C-band SAR data for wind speed retrieval, Zhang et al. (2011) have shown that in contrast to co-polarised data, the NRCS of cross-polarised data does not saturate under strong wind conditions and is further not sensitive to incidence angles or wind directions. This finding suggests that under high winds speeds, the elevated and non-saturated backscatter signature of VH-polarised data might lead to a very low land-water contrast, potentially resulting in a higher amount of misclassifications. In the same manner, VV polarised data might suffer from a high backscatter variability of open water surfaces when wind directions vary across a given satellite scene. While HH polarisation is generally considered as superior to VV or VH polarisation for flood mapping purposes (Henry et al. 2008) since it yields the highest contrast between open water and land areas, this polarisation is usually not available for systematically acquired Sentinel-1 data of land surfaces.
The computational effort of the complete workflow was approx. 45 min. for a Sentinel-1 IW mode GRD scene using an Intel Xeon E5-4650 CPU (8 cores) with 2.7 GHz and 16 GB of RAM on a Linux-based 64-bit operating system.
Outlook and further development
In a few critical situations the global threshold derived from the tile-based automatic thresholding algorithm may not be optimal due to following considerations:
• The SAR signal from C-band, the working frequency of Sentinel-1 SAR sensor, strongly attenuates vegetated areas, which also results in a dominant contrast similar to that generated by surface water.
• The quality of selected tiles and their derived thresholds are not evaluated in the current implementation. Due to that overestimations might occur in rare cases.
In order to solve the abovementioned problems, a hierarchical version of KI algorithm was implemented, inspired and adapted from Chini et al. (2016). The original idea of Chini et al. (2016) is to transform a SAR image into a quad-tree of small regular split tiles, and then check the bimodality of the histogram for each tile using a curve fitting technique, later those tiles with a bimodal histogram will be collected and merged. This procedure starts from the root of the quad-tree (coarsest resolution) and repeats to its bottom (finest resolution). Finally all collected tiles at the finest resolution level (30² pixel) will be merged as a bimodal subset image, which is used to derive a global threshold by KI. The advantage of this implementation is that every selected tile has been checked for the bimodality of its histogram, so the final merged subset image is also supposed to have a bimodal histogram. However, in order to get a merged bimodal subset, this procedure have
This can impair the efficiency of the program. One difference between our current implementation to that by Chini et al. (2016) is that we do not try to collect and use all tiles with bimodal histograms to form a global subset image. Instead, at first we rank and sort all split tiles according to a contrast measure and save the result as a list. In this way we know that the most appropriate representatives for threshold estimation lie at the top of the list. The bimodality of each tile in that list is checked from top to bottom until enough representatives (current number is 5 or 7) are received. In this way we reduce the amount of data to be processed but still attain a robust result.
During the procedure of tile selection and threshold estimation, the program starts with tiles that have a strong contrast and a large size. In figure 2.1.12, a candidate tile is presented which is about 532 x 362 pixel in dimension and has the strongest contrast, its histogram is illustrated on the right hand side which shows an ambiguous bimodal curve. The derived threshold is -15.9 dB which is overestimated and rejected due to a relative low Ashman’s separability of the histogram (about 3.38).
Figure 2.1.12: A rejected tile with a strong contrast but a poor bimodal histogram.
Figure 2.1.13: An accepted tile with a strong contrast and a pronounced bimodal histogram.
An accepted tile is illustrated in figure 2.1.13 which size is 266 x180 pixels. The Ashman’s separability of the histogram is about 3.98 which demonstrates a clear bimodality. The result of threshold estimation is -22.7 dB, which is an improved threshold value to separate dark surface water from vegetated background areas with higher backscatter.
2.1.2.2 Change detection
Change detection is the process of identifying differences in the state of an object or phenomenon by observing it at different times (Lu et al., 2004). From the viewpoint of crisis monitoring these changes observed from remote sensing data may be significantly related to natural disaster events and are therefore of huge concern for civil security. Along with SAR which enables all-weather data acquisition change detection provides an effective framework to detect NRT changes caused by natural disaster and extract relevant crisis information. In the following, change detection methodologies geared towards a) the visualisation of changes based on RGB composites as well as b) the classification of changes will be analysed and discussed.
Change visualisation based on RGB colour composites
Besides change detection methodologies used to classify areas of change, which is subject of a separate sub-chapter, it is often desirable to only visualise changes for dissemination purposes or to facilitate thematic analysis by an image interpreter. For the purpose of visualising backscatter changes due to flood situations, a relatively simple but effective methodology has been developed. The technique is based on the combination of pre- and post-flood SAR imagery into a RGB composite which is intuitively interpretable. Backscatter changes between two given image dates are scaled and colour-coded in a way to highlight flooded areas in blueish tones, positive backscatter changes (e.g. due to land cover transition) in reddish tones, areas with open water surfaces on both image dates in black colour and relatively unchanged areas in greyish tones (fig. 2.1.14). In the following, the methodology for SAR-based RGB-visualisation of flood-related changes is outlined and demonstrated.
Blue channel: The blue channel principally controls the colour coding of negative backscatter changes to blueish tones in the change RGB. This is performed through the following steps:
• Calculation of the normalised change index (NCI) between pre- and post-flood scenes.
• Condition A: Use of the layer in case a) its pixel value is below the mean of the NCI-layer minus its standard deviation and if b) the corresponding pixel value of the pre-disaster image is above a user-defined upper backscatter threshold for open water areas.
• Condition B: If condition A is not fulfilled, the average pixel value of pre- and post-disaster scenes is employed for the blue channel.
Green channel: The backscatter information of the pre-disaster scene is utilised.
Figure 2.1.14: RGB colour composite based on pre- and post-flood TerraSAR-X data. Blue: areas of negative backscatter changes related to floods, Red/magenta: Positive backscatter changes, e.g. related to land cover transition, black: areas with open water surfaces on both image dates, grey: land surfaces with no or minor backscatter changes. Pre-disaster scene: TerraSAR-X SM, HH, 28-12-2010, Post-disaster scene: TerraSAR-X SM, HH, 2011-01-08.
The methodology is principally applicable to all SAR data. In order to minimise changes due to other factors than flood events (e.g. seasonal changes or land cover transition), the time-span between pre- and post-flood acquisition should be kept as short as possible.
In figure 2.1.15, the results for two Sentinel-1 scenes of the 2015 flood event in Malawi (see chapter 2.1.2) are shown. While the flood extent is still very well perceptible in blueish tones, also many changes unrelated to the flood event can be observed which can be attributed to the longer time-span between pre- and post-flood acquisitions (24 days) and the land cover types prevalent in this test site. In direct vicinity to the flood extent, areas with strong reddish tones are visible, indicating a strong backscatter increase. This backscatter increase can potentially be attributed to double bounce effects from partly submerged vegetation.
Figure 2.1.15: RGB colour composite based on pre- and post-flood Sentinel-1 data. Blue: areas of negative backscatter changes related to floods, Red/magenta: Positive backscatter changes, e.g. related to double bounce effects from partially submerged vegetation, black: areas with open water surfaces on both image dates. Pre-disaster scene: Sentinel-1 IW, VV, 2015-01-04; post-disaster scene: Sentinel-1 IW, VV, 2015-01-28.
Change detection based on automatic thresholding and graph cut algorithms
Related works in the area of change detection can be found in reviews given by Lu et al. (2004) and Radke et al. (2005). According to their description change detection techniques can be divided into following groups:
1. Algebraic methods such as image differencing, rationing and index differencing. 2. Transformation methods like principal component analysis (PCA).
3. Classification methods for multi-spectral images such as spectral-temporal combined hybrid change detection and artificial neural network (ANN).
4. Advanced models, for example a biophysical model related to the scattering process of vegetation to find changed areas.
In order to provide NRT results only time-efficient methods should be taken into account, therefore methods from the first group are preferred. At the same time an additional post-processing method is always necessary due to the speckle noise of SAR data. In this way we decide to apply a hybrid method that combines an efficient thresholding method with a post-processing algorithm from computer vision. Our method is designed as follows:
1. Based on the statistics of the ratio image the optimal thresholds for negative- and positive-change classes are derived automatically. This step provides at first an initial image classification.
2. The initial classification is typically noisy. This noise signal on SAR images can be modelled as a Markov Random Field (MRF) (Li 1995) which can be optimised using the graph-cut method by Boykov et al. (2011) and Kolmogorov et al. (2004).
In the next paragraphs we describe an automatic workflow of the proposed SAR-based automatic three-class change detection processor shown in figure 2.1.16, which consists of three sub-chains: a) a pre-processing step for radiometric and geometrical calibration of input Sentinel-1 GRD SAR data, b) an initial classification based on a threshold derived by hypotheses test, and c) a post classification using a graph-cut solver for a general MRF model. More details of each sub-chain are given in the following subsections.
Figure 2.1.16: Workflow of the proposed SAR based automatic three-class change detection processor.
Geometric and radiometric calibration
The two input Sentinel-1 GRD SAR data sets are at first radiometrically calibrated to sigma naught (dB) using an included calibration Look Up Table (LUT), followed by a geocoding to GCS84. A Range-Doppler terrain correction is additionally integrated to correct geometric distortions of GRD products due to terrain effects. The whole calibration is implemented using the Graph Processing Tool (GPT) of ESA Sentinel-1 Toolbox (S1TBX), which is a command-line interface easily reusable in other programming languages, such as IDL used in this study. In addition, Shuttle Radar Topography Mission (SRTM) 1 arc second tiles corresponding to the given Sentinel-1 scenes are automatically downloaded by S1TBX for terrain correction. Subsequently, one of the two calibrated Sentinel-1 GRD SAR scenes is clipped and resampled in IDL to ensure the same image extent and
